Re: Thinking of using ZFS/FBSD for a backup system

Dmitry Morozovsky wrote:
> On Wed, 9 Jul 2008, Kris Kennaway wrote:
> 
> KK> > marck_at_wizzle:/usr/ports> grep -Ril zfs Tools/
> KK> > Tools/portbuild/scripts/claim-chroot
> KK> > Tools/portbuild/scripts/clean-chroot
> KK> > Tools/portbuild/scripts/cleanup-chroots
> KK> > 
> KK> > ;-)
> KK> > 
> KK> > Any pitfalls while using this? Thanks!
> KK> 
> KK> It's still settling on pointyhat, but you could check it out there if you
> KK> like.  The most annoying thing is actually a limitation of the FreeBSD NFS
> KK> server, since it returns errors to clients while it is reloading the export
> KK> list.
> 
> Argh. And, which process is reporting an error? nfsd or some of kthreads? In 
> the former case, is it possible to create new master-slave process 
> tree, rebinding and signalling old tree to exit when all requests are finished?
> 
> The same approach is used by www/nginx to both seamless configuration change 
> and also binary file upgrade.
> 
> CC:ing -current_at_ to ensure the theme have a bit broader audience.

The problem is that updating the NFS export list is not atomic. 
Instead, it is deleted and reloaded.  This means that client I/O 
requests that are received during the window do not correspond to a 
valid export, and the server dutifully returns an error to the client.

This is made worse by the practice of mount(8) (and zfs(8)) of 
SIGHUP'ing mountd every time a new filesystem is mounted (even if it's 
not exported).  I have disabled this locally, but it doesn't help when I 
really do want to export a newly mounted filesystem but I have clients 
doing I/O.

A few years ago there was a patch from Andrey Simonenko 
<simon_at_comsys.ntu-kpi.kiev.ua> that implemented atomically reloading the 
mount list, although it was unfortunately a mixture of these and other 
changes including some that might be seen as gratuitous (e.g. it 
depended on a new library).  Unfortunately the patch no longer seems to 
be available on the original website and my attempt to contact Andrey 
has failed.

Kris